whispercpp 1.3.6 → 1.3.8

data/ext/sources/ggml/src/ggml-openvino/openvino/op/rope.cpp

@@ -7,14 +7,20 @@
 #include <openvino/core/node.hpp>
 #include <openvino/core/node_output.hpp>
 #include <openvino/op/add.hpp>
+#include <openvino/op/broadcast.hpp>
 #include <openvino/op/concat.hpp>
 #include <openvino/op/constant.hpp>
+#include <openvino/op/convert.hpp>
+#include <openvino/op/cos.hpp>
+#include <openvino/op/gather.hpp>
 #include <openvino/op/multiply.hpp>
 #include <openvino/op/reshape.hpp>
 #include <openvino/op/shape_of.hpp>
+#include <openvino/op/sin.hpp>
 #include <openvino/op/slice.hpp>
 #include <openvino/op/split.hpp>
 #include <openvino/op/subtract.hpp>
+#include <openvino/op/transpose.hpp>
 #include <openvino/op/unsqueeze.hpp>
 #include <vector>
 
@@ -33,6 +39,11 @@ OutputVector translate_rope(const NodeContext & context) {
     auto data_node = context.get_input(0).get_node_shared_ptr();
     auto output_shape = context.get_output_shape().to_shape();
     int32_t * op_params = context.get_output_op_params();
+    const int mode = op_case;
+
+    constexpr int TYPE_NORMAL = 0;
+    constexpr int TYPE_NEOX = 1;
+    constexpr int TYPE_IMROPE = 2;
 
     Output<Node> cos_theta_node;
     Output<Node> sin_theta_node;
@@ -45,59 +56,146 @@ OutputVector translate_rope(const NodeContext & context) {
         if (context.get_input_size() == 3) {
             rope_freqs_weight = context.get_input(2).get_node_shared_ptr();
         }
-        auto sin_cos = make_sin_cos(op_params, inp_pos, rope_freqs_weight);
+        auto sin_cos = make_sin_cos(op_params, inp_pos, rope_freqs_weight, mode == TYPE_IMROPE, false);
         sin_theta_node = sin_cos.first;
         cos_theta_node = sin_cos.second;
     }
 
-    if (op_case == 2) {
-        // The input comes from a VIEW
-        int slice_len = output_shape[2] * output_shape[3];
-        data_node = process_view_input(context, 0, slice_len).get_node_shared_ptr();
+    if (context.get_view_input_size(0) > 0) {
+        data_node = process_view_input_new(context, 0).get_node_shared_ptr();
         if (context.is_stateful()) {
             auto data_shape = ov::op::v0::Constant::create(
                 ov::element::i64, {3}, std::vector<int64_t>{-1, (int64_t) output_shape[2], (int64_t) output_shape[3]});
             data_node = std::make_shared<ov::op::v1::Reshape>(data_node, data_shape, false);
         } else {
             auto data_shape = ov::op::v0::Constant::create(
-                ov::element::i64, {4}, std::vector<int64_t>{1, -1, (int64_t) output_shape[2], (int64_t) output_shape[3]});
+                ov::element::i64, {4},
+                std::vector<int64_t>{1, -1, (int64_t) output_shape[2], (int64_t) output_shape[3]});
             data_node = std::make_shared<ov::op::v1::Reshape>(data_node, data_shape, false);
         }
     }
 
-    const int mode = op_params[2];
-    constexpr int ROPE_TYPE_NORMAL = 0;
-    constexpr int ROPE_TYPE_NEOX = 2;
-
-    if (mode == ROPE_TYPE_NORMAL) {
-        auto neg_one = ov::op::v0::Constant::create(ov::element::i64, {1}, {-1});
-        auto zero = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
-        auto one = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
-        auto two = ov::op::v0::Constant::create(ov::element::i64, {1}, {2});
-        auto end = ov::op::v0::Constant::create(ov::element::i64, {1}, {output_shape[3]});
-        Output<Node> even_slice;
-        Output<Node> odd_slice;
-        int32_t unsqueeze_dim = context.is_stateful() ? 3 : 4;
-        even_slice = std::make_shared<ov::op::v8::Slice>(data_node, zero, end, two, neg_one);
-        odd_slice = std::make_shared<ov::op::v8::Slice>(data_node, one, end, two, neg_one);
-
-        Output<Node> first_half =
-            std::make_shared<ov::op::v1::Subtract>(std::make_shared<ov::op::v1::Multiply>(even_slice, cos_theta_node),
-                                                   std::make_shared<ov::op::v1::Multiply>(odd_slice, sin_theta_node));
-        Output<Node> second_half =
-            std::make_shared<ov::op::v1::Add>(std::make_shared<ov::op::v1::Multiply>(even_slice, sin_theta_node),
-                                              std::make_shared<ov::op::v1::Multiply>(odd_slice, cos_theta_node));
-
-        first_half = std::make_shared<ov::op::v0::Unsqueeze>(first_half,
-                                                             ov::op::v0::Constant::create(ov::element::i64, {1}, {unsqueeze_dim}));
-        second_half = std::make_shared<ov::op::v0::Unsqueeze>(second_half,
-                                                              ov::op::v0::Constant::create(ov::element::i64, {1}, {unsqueeze_dim}));
-        auto stack = std::make_shared<ov::op::v0::Concat>(OutputVector{first_half, second_half}, unsqueeze_dim);
-
-        auto data_shape = ov::op::v0::Constant::create(
-            ov::element::i64, {4}, std::vector<int64_t>{1, -1, (int64_t) output_shape[2], (int64_t) output_shape[3]});
-        res = std::make_shared<ov::op::v1::Reshape>(stack, data_shape, false);
-    } else if (mode == ROPE_TYPE_NEOX) {
+    auto output_type = context.get_output_type();
+    if (data_node->get_element_type() != ov::element::f32) {
+        data_node = std::make_shared<ov::op::v0::Convert>(data_node, ov::element::f32);
+    }
+
+    // TODO(openvino-gpu-rope-fusion): TEMPORARY WORKAROUND - do NOT revert until the
+    // OpenVINO GPU plugin is updated.
+    //
+    // For TYPE_NORMAL rope (both stateful and stateless) we emit the Flux-style
+    // interleaved pattern below so the GPU plugin's RoPEFusionFlux matcher folds it
+    // into ov::op::internal::RoPE. The matcher requires rank-4 inputs, which is why
+    // the original even/odd Slice translation (kept in the `else if (mode ==
+    // TYPE_NORMAL)` branch below for reference) does not get fused.
+    //
+    // Once the GPU plugin's RoPE fusion is extended to also recognize the original
+    // even/odd Slice form, this Flux rewrite should be removed and both modes should
+    // be restored to the captured even/odd translation. Until then, keep both paths:
+    // the active Flux rewrite here and the previous translation preserved below.
+    if (mode == TYPE_NORMAL) {
+        // Emit the Flux-style interleaved-RoPE pattern so the GPU plugin's
+        // RoPEFusionFlux matcher folds this subgraph into ov::op::internal::RoPE:
+        //   x_paired   = Reshape(x, [1, S, n_heads, head_size/2, 2])
+        //   x0, x1     = Split(x_paired, axis=-1, num_splits=2)
+        //   x1_neg     = x1 * -1
+        //   x_rotated  = Reshape(Concat([x1_neg, x0], axis=-1), [1, S, n_heads, head_size])
+        //   y          = x * t_cos + x_rotated * t_sin
+        // Mathematically equivalent to the even/odd Slice form below.
+        //
+        // RoPEFusionFlux requires rank_equals(4) on x, t_cos and t_sin. The cos/sin
+        // tables are already built rank-4 ([1, S, 1, head_size/2]) for both modes. In
+        // stateful mode the data arrives rank-3 ([S, n_heads, head_size]), so lift it
+        // to rank-4 ([1, S, n_heads, head_size]) here. Stateful RoPE already produced
+        // rank-4 output, so downstream attention is unaffected.
+        if (context.is_stateful()) {
+            auto r4_shape = ov::op::v0::Constant::create(
+                ov::element::i64, {4},
+                std::vector<int64_t>{1, -1, (int64_t) output_shape[2], (int64_t) output_shape[3]});
+            data_node = std::make_shared<ov::op::v1::Reshape>(data_node, r4_shape, false);
+        }
+        const int64_t head_size = static_cast<int64_t>(output_shape[3]);
+        const int64_t n_heads = static_cast<int64_t>(output_shape[2]);
+        const int64_t half = head_size / 2;
+
+        auto neg_one_f = ov::op::v0::Constant::create(data_node->get_element_type(), ov::Shape{}, {-1.0f});
+
+        auto paired_shape =
+            ov::op::v0::Constant::create(ov::element::i64, {5}, std::vector<int64_t>{1, -1, n_heads, half, 2});
+        auto x_paired = std::make_shared<ov::op::v1::Reshape>(data_node, paired_shape, false);
+
+        auto split_axis = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{}, {-1});
+        auto data_split = std::make_shared<ov::op::v1::Split>(x_paired, split_axis, 2);
+        Output<Node> x0 = data_split->outputs()[0];
+        Output<Node> x1 = data_split->outputs()[1];
+
+        auto x1_neg = std::make_shared<ov::op::v1::Multiply>(x1, neg_one_f);
+        auto x_rotated_paired = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{x1_neg, x0}, -1);
+
+        auto flat_shape =
+            ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{1, -1, n_heads, head_size});
+        auto x_rotated = std::make_shared<ov::op::v1::Reshape>(x_rotated_paired, flat_shape, false);
+
+        // Expand cos/sin from [..., head_size/2] to [..., head_size] by repeating each
+        // entry twice. Use special_zero on the final Reshape so the seq dim passes
+        // through dynamically. Final rank is 4 to satisfy the matcher's predicate.
+        auto expand_cos_sin = [&](Output<Node> cs) {
+            auto cs_unsq =
+                std::make_shared<ov::op::v0::Unsqueeze>(cs, ov::op::v0::Constant::create(ov::element::i64, {1}, {-1}));
+            auto bcast_target =
+                ov::op::v0::Constant::create(ov::element::i64, {5}, std::vector<int64_t>{1, 1, 1, half, 2});
+            auto bcast =
+                std::make_shared<ov::op::v3::Broadcast>(cs_unsq, bcast_target, ov::op::BroadcastType::BIDIRECTIONAL);
+            auto flat = ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{0, 0, 0, head_size});
+            return std::make_shared<ov::op::v1::Reshape>(bcast, flat, true);
+        };
+        Output<Node> cos_full = expand_cos_sin(cos_theta_node);
+        Output<Node> sin_full = expand_cos_sin(sin_theta_node);
+
+        auto y1 = std::make_shared<ov::op::v1::Multiply>(data_node, cos_full);
+        auto y2 = std::make_shared<ov::op::v1::Multiply>(x_rotated, sin_full);
+        res = std::make_shared<ov::op::v1::Add>(y1, y2);
+    }
+    // PRESERVED PREVIOUS TRANSLATION - Re-enable this branch (and remove the Flux branch above) once
+    // the GPU plugin's RoPE fusion is updated to recognize the even/odd Slice form;
+    // see the TODO(openvino-gpu-rope-fusion) note above. Do not delete.
+    //
+    // Original even/odd Slice form. In stateless mode it ran on rank-4 data
+    // ([1, S, n_heads, head_size]); in stateful mode on rank-3 data
+    // ([S, n_heads, head_size]). Either way it does not match RoPEFusionFlux
+    // (which needs rank-4 x in the interleaved layout), so the RoPE stays as
+    // discrete elementwise ops.
+    //
+    // } else if (mode == TYPE_NORMAL) {
+    //     auto neg_one = ov::op::v0::Constant::create(ov::element::i64, {1}, {-1});
+    //     auto zero = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
+    //     auto one = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+    //     auto two = ov::op::v0::Constant::create(ov::element::i64, {1}, {2});
+    //     auto end = ov::op::v0::Constant::create(ov::element::i64, {1}, {output_shape[3]});
+    //     Output<Node> even_slice;
+    //     Output<Node> odd_slice;
+    //     // stateful data is rank 3 (unsqueeze at axis 3), stateless is rank 4 (axis 4)
+    //     int32_t unsqueeze_dim = context.is_stateful() ? 3 : 4;
+    //     even_slice = std::make_shared<ov::op::v8::Slice>(data_node, zero, end, two, neg_one);
+    //     odd_slice = std::make_shared<ov::op::v8::Slice>(data_node, one, end, two, neg_one);
+    //
+    //     Output<Node> first_half =
+    //         std::make_shared<ov::op::v1::Subtract>(std::make_shared<ov::op::v1::Multiply>(even_slice, cos_theta_node),
+    //                                                std::make_shared<ov::op::v1::Multiply>(odd_slice, sin_theta_node));
+    //     Output<Node> second_half =
+    //         std::make_shared<ov::op::v1::Add>(std::make_shared<ov::op::v1::Multiply>(even_slice, sin_theta_node),
+    //                                           std::make_shared<ov::op::v1::Multiply>(odd_slice, cos_theta_node));
+    //
+    //     first_half = std::make_shared<ov::op::v0::Unsqueeze>(first_half,
+    //                                                          ov::op::v0::Constant::create(ov::element::i64, {1}, {unsqueeze_dim}));
+    //     second_half = std::make_shared<ov::op::v0::Unsqueeze>(second_half,
+    //                                                           ov::op::v0::Constant::create(ov::element::i64, {1}, {unsqueeze_dim}));
+    //     auto stack = std::make_shared<ov::op::v0::Concat>(OutputVector{first_half, second_half}, unsqueeze_dim);
+    //
+    //     auto data_shape = ov::op::v0::Constant::create(
+    //         ov::element::i64, {4}, std::vector<int64_t>{1, -1, (int64_t) output_shape[2], (int64_t) output_shape[3]});
+    //     res = std::make_shared<ov::op::v1::Reshape>(stack, data_shape, false);
+    else if (mode == TYPE_NEOX) {
         auto data_split = std::make_shared<ov::op::v1::Split>(
             data_node, ov::op::v0::Constant::create(ov::element::i64, ov::Shape{}, {-1}), 2);
         Output<Node> slice_data_node_0 = data_split->outputs()[0];
@@ -112,6 +210,30 @@ OutputVector translate_rope(const NodeContext & context) {
             std::make_shared<ov::op::v1::Multiply>(slice_data_node_1, cos_theta_node));
 
         res = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{first_half_node, second_half_node}, -1);
+    } else if (mode == TYPE_IMROPE) {
+        int64_t n_dims = data_node->get_output_partial_shape(0)[3].get_length();
+        auto cos_sin_shape = std::make_shared<ov::op::v0::Constant>(ov::element::i64, ov::Shape{4},
+                                                                    std::vector<int64_t>{1, -1, 1, (n_dims >> 1)});
+        auto cos_reshaped = std::make_shared<ov::op::v1::Reshape>(cos_theta_node, cos_sin_shape, true);
+        auto sin_reshaped = std::make_shared<ov::op::v1::Reshape>(sin_theta_node, cos_sin_shape, true);
+
+        auto split_axis = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{}, {3});
+        auto split_a = std::make_shared<ov::op::v1::Split>(data_node, split_axis, 2);
+        auto x0 = split_a->output(0);
+        auto x1 = split_a->output(1);
+        auto mul_a = std::make_shared<ov::op::v1::Multiply>(x0, cos_reshaped);
+        auto mul_b = std::make_shared<ov::op::v1::Multiply>(x1, sin_reshaped);
+        auto sub = std::make_shared<ov::op::v1::Subtract>(mul_a, mul_b);
+
+        auto mul_c = std::make_shared<ov::op::v1::Multiply>(x0, sin_reshaped);
+        auto mul_d = std::make_shared<ov::op::v1::Multiply>(x1, cos_reshaped);
+        auto add = std::make_shared<ov::op::v1::Add>(mul_c, mul_d);
+
+        res = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{sub, add}, 3);
+    }
+
+    if (res.get_element_type() != output_type) {
+        res = std::make_shared<ov::op::v0::Convert>(res, output_type);
     }
 
     return rename_outputs_with_suffix({res}, context.get_name());

data/ext/sources/ggml/src/ggml-openvino/openvino/op/set_rows.cpp

@@ -28,20 +28,20 @@ namespace op {
 OutputVector translate_set_rows(const NodeContext & context) {
     num_inputs_check(context, 3, 3);
 
-    auto data = context.get_input(0);
+    auto data = process_view_input_new(context, 0);
     auto indices = context.get_input(1);
     auto dst = context.get_input(2);
 
     data = std::make_shared<ov::op::v0::Convert>(data, context.get_output_type());
 
-    auto dst_shape = context.get_output_shape().to_shape();
+    auto row_size = context.get_input_shape(2)[3].get_length();
 
     auto ind_squeezed =
         std::make_shared<ov::op::v0::Squeeze>(indices, ov::op::v0::Constant::create(ov::element::i64, {3}, {0, 1, 2}));
     auto data_reshaped = std::make_shared<ov::op::v1::Reshape>(
         data,
         ov::op::v0::Constant::create(ov::element::i64, {4},
-                                     {(int64_t) 1, (int64_t) 1, (int64_t) -1, (int64_t) dst_shape[3]}),
+                                     {(int64_t) 1, (int64_t) 1, (int64_t) -1, (int64_t) row_size}),
         false);
     auto axes = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{}, {2});
 

data/ext/sources/ggml/src/ggml-openvino/openvino/op/softmax.cpp

@@ -2,17 +2,19 @@
 #include "../op_table.h"
 #include "../utils.h"
 
-#include <climits>
+#include <cmath>
 #include <cstdint>
+#include <cstring>
 #include <memory>
-#include <openvino/core/node.hpp>
-#include <openvino/core/node_output.hpp>
+#include <openvino/op/broadcast.hpp>
+#include <openvino/frontend/exception.hpp>
 #include <openvino/op/add.hpp>
 #include <openvino/op/concat.hpp>
 #include <openvino/op/constant.hpp>
 #include <openvino/op/convert.hpp>
-#include <openvino/op/matmul.hpp>
 #include <openvino/op/multiply.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/shape_of.hpp>
 #include <openvino/op/slice.hpp>
 #include <openvino/op/softmax.hpp>
 #include <vector>
@@ -22,63 +24,138 @@ namespace frontend {
 namespace ggml {
 namespace op {
 
-OutputVector translate_soft_max(const NodeContext & context) {
-    // TODO code is outdated
-    num_inputs_check(context, 1, 2);
+static bool is_static_one(const ov::Dimension & dim) {
+    return dim.is_static() && dim.get_length() == 1;
+}
+
+static bool same_static_dim(const ov::Dimension & lhs, const ov::Dimension & rhs) {
+    return lhs.is_static() && rhs.is_static() && lhs.get_length() == rhs.get_length();
+}
+
+static bool is_attention_sinks_input_shape(const ov::PartialShape & candidate, const ov::PartialShape & logits_shape) {
+    if (candidate.rank().is_dynamic() || logits_shape.rank().is_dynamic() || candidate.rank().get_length() != 4 ||
+        logits_shape.rank().get_length() != 4) {
+        return false;
+    }
+
+    return is_static_one(candidate[0]) && is_static_one(candidate[1]) && is_static_one(candidate[2]) &&
+           same_static_dim(candidate[3], logits_shape[1]);
+}
 
-    auto input_node = context.get_input(0).get_node_shared_ptr();
-    ov::Output<Node> res;
+// Reimplementation of GGML_OP_SOFT_MAX semantics for OpenVINO backend:
+// 1) logits = src0 * scale
+// 2) logits += mask (if provided)
+// 3) append attention sinks as hidden logits (if provided)
+// 4) softmax over the last dimension and remove the hidden sink column
+OutputVector translate_soft_max(const NodeContext & context) {
+    num_inputs_check(context, 1, 3);
 
     float scale = 1.0f;
     float max_bias = 0.0f;
-    auto * op_params = context.get_output_op_params();
-    memcpy(&scale, (float *) op_params + 0, sizeof(float));
-    memcpy(&max_bias, (float *) op_params + 1, sizeof(float));
-    auto src0_shape = context.get_input_shape(0).get_shape();
-    const uint32_t h = src0_shape[2];
-    const uint32_t n_head = src0_shape[0];
-    const uint32_t n_head_log2 = 1u << (uint32_t) floor(log2(n_head));
-
-    const float m0 = powf(2.0f, -(max_bias) / n_head_log2);
-    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
-    const float slope =
-        (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2 * (h - n_head_log2) + 1) : 1.0f;
-
-    auto scale_node = std::make_shared<ov::op::v0::Constant>(ov::element::f32, ov::Shape{}, std::vector<float>{scale});
-    auto scaled_input = std::make_shared<ov::op::v1::Multiply>(input_node, scale_node);
-
-    if (context.get_input_size() < 2) {
-        res = std::make_shared<ov::op::v8::Softmax>(scaled_input, 2);
-        return rename_outputs_with_suffix({res}, context.get_name());
+    memcpy(&scale, (float *) context.get_output_op_params() + 0, sizeof(float));
+    memcpy(&max_bias, (float *) context.get_output_op_params() + 1, sizeof(float));
+
+    ov::Output<ov::Node> logits = context.get_input(0);
+    const bool second_input_is_sinks =
+        context.get_input_size() == 2 && is_attention_sinks_input_shape(context.get_input_shape(1), context.get_output_shape());
+    const bool has_mask = context.get_input_size() > 1 && !second_input_is_sinks;
+    const bool has_sinks = second_input_is_sinks || context.get_input_size() > 2;
+    const size_t sinks_input_idx = second_input_is_sinks ? 1 : 2;
+
+    // Apply scale first: logits = src0 * scale
+    if (scale != 1.0f) {
+        auto scale_const =
+            std::make_shared<ov::op::v0::Constant>(ov::element::f32, ov::Shape{}, std::vector<float>{scale});
+        logits = std::make_shared<ov::op::v1::Multiply>(logits, scale_const);
     }
 
-    ov::Output<ov::Node> mask_node_sliced;
-    if (context.has_input("KQ_mask_sliced")) {
-        mask_node_sliced = context.get_input("KQ_mask_sliced");
-    } else {
-        auto token_len = get_dimensions(input_node, {1});
-        auto mask_node = context.get_input(1);
+    FRONT_END_CHECK_IMPLEMENTED(!(max_bias > 0.0f && !has_mask),
+                                "OpenVINO softmax ALiBi path requires mask input");
+
+    // Optional mask add: logits += mask
+    // For max_bias > 0 (ALiBi), apply per-head slope to mask before adding.
+    if (has_mask) {
+        ov::Output<ov::Node> mask = context.get_input(1);
+
+        // For stateful
+        std::string mask_name = "KQ_mask_sliced";
+        if (context.get_input_names()[1].find("swa") != std::string::npos) {
+            mask_name = "KQ_mask_swa_sliced";
+        }
+        if (context.has_input(mask_name)) {
+            mask = context.get_input(mask_name);
+        }
+
+        if (mask.get_element_type() != logits.get_element_type()) {
+            mask = std::make_shared<ov::op::v0::Convert>(mask, logits.get_element_type());
+        }
+
+        if (max_bias > 0.0f) {
+            auto out_shape = context.get_output_shape().to_shape();
+            FRONT_END_CHECK_IMPLEMENTED(out_shape.size() == 4, "OpenVINO softmax ALiBi path expects rank-4 tensor");
+
+            const uint32_t n_head = static_cast<uint32_t>(out_shape[1]);
+            FRONT_END_CHECK_IMPLEMENTED(n_head > 0, "OpenVINO softmax ALiBi path expects n_head > 0");
+
+            const uint32_t n_head_log2 = 1u << static_cast<uint32_t>(std::floor(std::log2(static_cast<float>(n_head))));
+            const float m0 = std::pow(2.0f, -(max_bias) / static_cast<float>(n_head_log2));
+            const float m1 = std::pow(2.0f, -(max_bias / 2.0f) / static_cast<float>(n_head_log2));
+
+            std::vector<float> slopes(n_head);
+            for (uint32_t h = 0; h < n_head; ++h) {
+                slopes[h] = h < n_head_log2 ? std::pow(m0, static_cast<float>(h + 1)) :
+                                              std::pow(m1, static_cast<float>(2 * (h - n_head_log2) + 1));
+            }
+
+            ov::Output<ov::Node> slope_node =
+                std::make_shared<ov::op::v0::Constant>(ov::element::f32, ov::Shape{n_head}, slopes);
+            if (slope_node.get_element_type() != mask.get_element_type()) {
+                slope_node = std::make_shared<ov::op::v0::Convert>(slope_node, mask.get_element_type());
+            }
+
+            auto slope_shape = std::make_shared<ov::op::v0::Constant>(
+                ov::element::i64, ov::Shape{4}, std::vector<int64_t>{1, static_cast<int64_t>(n_head), 1, 1});
+            auto slope_4d = std::make_shared<ov::op::v1::Reshape>(slope_node, slope_shape, false);
+            mask = std::make_shared<ov::op::v1::Multiply>(mask, slope_4d);
+        }
+
+        logits = std::make_shared<ov::op::v1::Add>(logits, mask);
+    }
+
+    ov::Output<ov::Node> softmax_input = logits;
+    if (has_sinks) {
+        ov::Output<ov::Node> sinks = context.get_input(sinks_input_idx);
+        if (sinks.get_element_type() != logits.get_element_type()) {
+            sinks = std::make_shared<ov::op::v0::Convert>(sinks, logits.get_element_type());
+        }
+
+        auto sink_shape = ov::op::v0::Constant::create(ov::element::i64, {4}, {1, -1, 1, 1});
+        auto sinks_4d = std::make_shared<ov::op::v1::Reshape>(sinks, sink_shape, false);
+
+        auto logits_shape = std::make_shared<ov::op::v3::ShapeOf>(logits, ov::element::i64);
         auto zero = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
         auto one = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
-        mask_node_sliced = std::make_shared<ov::op::v8::Slice>(mask_node, zero, token_len, one, one);
-    }
+        auto three = ov::op::v0::Constant::create(ov::element::i64, {1}, {3});
+        auto four = ov::op::v0::Constant::create(ov::element::i64, {1}, {4});
+        auto shape_axis = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
 
-    if (mask_node_sliced.get_element_type() != context.get_output_type()) {
-        mask_node_sliced = std::make_shared<ov::op::v0::Convert>(mask_node_sliced, context.get_output_type());
-    }
+        auto sink_prefix_shape = std::make_shared<ov::op::v8::Slice>(logits_shape, zero, three, one, shape_axis);
+        auto sink_last_dim = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+        auto sink_broadcast_shape = std::make_shared<ov::op::v0::Concat>(
+            ov::OutputVector{sink_prefix_shape, sink_last_dim}, 0);
+        auto sink_column = std::make_shared<ov::op::v3::Broadcast>(sinks_4d, sink_broadcast_shape,
+                                                                   ov::op::BroadcastType::BIDIRECTIONAL);
+        softmax_input = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{logits, sink_column}, 3);
 
-    Output<Node> slope_mask;
-    if (slope != 1.0f) {
-        auto slope_node =
-            std::make_shared<ov::op::v0::Constant>(ov::element::f32, ov::Shape{}, std::vector<float>{slope});
-        slope_mask = std::make_shared<ov::op::v1::Multiply>(mask_node_sliced, slope_node);
-        throw std::runtime_error("Slope != 1.0f in softmax has not been tested, verify it before use.");
-    }
-    slope_mask = mask_node_sliced;
+        auto softmax_with_sink = std::make_shared<ov::op::v8::Softmax>(softmax_input, -1);
+        auto original_last_dim = std::make_shared<ov::op::v8::Slice>(logits_shape, three, four, one, shape_axis);
+        auto res = std::make_shared<ov::op::v8::Slice>(softmax_with_sink, zero, original_last_dim, one, three);
 
-    auto input_slope_mask_node = std::make_shared<ov::op::v1::Add>(scaled_input, slope_mask);
+        return rename_outputs_with_suffix({res}, context.get_name());
+    }
 
-    res = std::make_shared<ov::op::v8::Softmax>(input_slope_mask_node, 2);
+    // Softmax along last dimension (equivalent to ggml softmax over ne[0]).
+    auto res = std::make_shared<ov::op::v8::Softmax>(softmax_input, -1);
 
     return rename_outputs_with_suffix({res}, context.get_name());
 }

data/ext/sources/ggml/src/ggml-openvino/openvino/op/ssm_conv.cpp

@@ -0,0 +1,59 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <openvino/op/constant.hpp>
+#include <openvino/op/group_conv.hpp>
+#include <openvino/op/reshape.hpp>
+#include <openvino/op/transpose.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_ssm_conv(const NodeContext & context) {
+    num_inputs_check(context, 2, 2);
+
+    auto sx = context.get_input(0);                         // conv state + input: OV shape [1, n_s, d_inner, ncs]
+    auto c = context.get_input(1);                          // conv1d weight:      OV shape [1, 1, d_inner, d_conv]
+
+    auto sx_shape = context.get_input_shape(0).to_shape();  // [1, n_s, d_inner, ncs]
+    auto c_shape = context.get_input_shape(1).to_shape();   // [1, 1, d_inner, d_conv]
+
+    int64_t n_s = sx_shape[1];
+    int64_t d_inner = sx_shape[2];
+    int64_t ncs = sx_shape[3];  // d_conv - 1 + n_t
+    int64_t d_conv = c_shape[3];
+    int64_t n_t = ncs - d_conv + 1;
+
+    // Reshape sx from [1, n_s, d_inner, ncs] to [n_s, d_inner, ncs] for 1D GroupConvolution
+    auto sx_new_shape = ov::op::v0::Constant::create(ov::element::i64, {3}, std::vector<int64_t>{n_s, d_inner, ncs});
+    auto sx_reshaped = std::make_shared<ov::op::v1::Reshape>(sx, sx_new_shape, false);
+
+    // Reshape c from [1, 1, d_inner, d_conv] to [d_inner, 1, 1, d_conv]
+    // GroupConvolution filter: [groups, out_channels/groups, in_channels/groups, kernel_size]
+    auto c_new_shape = ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{d_inner, 1, 1, d_conv});
+    auto c_reshaped = std::make_shared<ov::op::v1::Reshape>(c, c_new_shape, false);
+
+    // Depthwise 1D convolution: groups=d_inner, stride=1, no padding, no dilation
+    // Input: [n_s, d_inner, ncs], Filter: [d_inner, 1, 1, d_conv]
+    // Output: [n_s, d_inner, n_t]
+    auto conv = std::make_shared<ov::op::v1::GroupConvolution>(
+        sx_reshaped, c_reshaped, ov::Strides{1}, ov::CoordinateDiff{0}, ov::CoordinateDiff{0}, ov::Strides{1});
+
+    // Transpose from [n_s, d_inner, n_t] to [n_s, n_t, d_inner]
+    auto perm = ov::op::v0::Constant::create(ov::element::i64, {3}, std::vector<int64_t>{0, 2, 1});
+    auto transposed = std::make_shared<ov::op::v1::Transpose>(conv, perm);
+
+    // Reshape to output shape [1, n_s, n_t, d_inner]
+    auto out_shape = ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{1, n_s, n_t, d_inner});
+    auto res = std::make_shared<ov::op::v1::Reshape>(transposed, out_shape, false);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

data/ext/sources/ggml/src/ggml-openvino/openvino/op/sum_rows.cpp

@@ -0,0 +1,27 @@
+#include "../node_context.h"
+#include "../op_table.h"
+#include "../utils.h"
+
+#include <memory>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/reduce_sum.hpp>
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace op {
+
+OutputVector translate_sum_rows(const NodeContext & context) {
+    num_inputs_check(context, 1, 1);
+
+    auto input = process_view_input_new(context, 0);
+    auto res = std::make_shared<ov::op::v1::ReduceSum>(
+        input, ov::op::v0::Constant::create(ov::element::i64, ov::Shape{1}, {-1}), true);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
+}  // namespace op
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

data/ext/sources/ggml/src/ggml-openvino/openvino/op/transpose.cpp

@@ -12,8 +12,39 @@ namespace op {
 OutputVector translate_transpose(const NodeContext & context) {
     num_inputs_check(context, 1, 1);
 
+    // Compute permute order from input/output shape and stride information
+    // so it adapts to different input and output layouts.
+    auto input_shape = context.get_input_shape(0).to_shape();
+    auto input_stride = context.get_input_stride(0);
+    auto output_shape = context.get_output_shape().to_shape();
+    auto output_stride = context.get_output_stride();
+
+    // Compute permute order by matching output and input stride rankings.
+    // Build <stride, dim_index> pairs.
+    std::vector<std::pair<size_t, int>> output_stride_dims;
+    std::vector<std::pair<size_t, int>> input_stride_dims;
+
+    for (int i = 0; i < 4; ++i) {
+        output_stride_dims.push_back({output_stride[i], i});
+        input_stride_dims.push_back({input_stride[i], i});
+    }
+
+    // Sort by stride in descending order.
+    std::sort(output_stride_dims.rbegin(), output_stride_dims.rend());
+    std::sort(input_stride_dims.rbegin(), input_stride_dims.rend());
+
+    // Build permute order.
+    std::vector<int64_t> permute_order(4);
+    for (int i = 0; i < 4; ++i) {
+        int output_dim = output_stride_dims[i].second;
+        int input_dim = input_stride_dims[i].second;
+        permute_order[output_dim] = input_dim;
+    }
+
+    auto input = process_view_input_new(context, 0);
+
     auto res = std::make_shared<ov::op::v1::Transpose>(
-        context.get_input(0), ov::op::v0::Constant::create(ov::element::i64, {4}, {0, 1, 3, 2}));
+        input, ov::op::v0::Constant::create(ov::element::i64, {4}, permute_order));
     return rename_outputs_with_suffix({res}, context.get_name());
 }
 

data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_silu.cpp

@@ -14,7 +14,7 @@ namespace op {
 OutputVector translate_unary_silu(const NodeContext & context) {
     num_inputs_check(context, 1, 1);
 
-    auto input = context.get_input(0);
+    auto input = process_view_input_new(context, 0);
     auto sigmoid = std::make_shared<ov::op::v0::Sigmoid>(input);
     auto res = std::make_shared<ov::op::v1::Multiply>(input, sigmoid);